Periodic Reporting for period 2 - MoP-MiP (Molecular Programming for MicroRNA profiling)
Reporting period: 2022-08-01 to 2024-01-31
MiRNAs, a class of RNA molecules responsible for the fine regulation of gene expression, are emerging as promising biomarkers for cancer monitoring (diagnostic, prognostic and prediction of therapy response). The measurement of miRNAs from liquid biopsies represents a hopeful opportunity to fight these diseases, enabling large population screening for early cancer appearance and real-time monitoring of the response to treatment. However, current miRNA detection methods do not meet the performances in terms of sensitivity, multiplexing and practicality, necessary to bring miRNA signatures in the front line of clinical applications.
MoP-MiP will develop novel approaches for miRNA profiling building on the latest advances in molecular programming (MP). MP deals with the design of artificial DNA reaction networks capable of information processing, which will be exploited throughout this project to build smarter biosensors. The team will pursue two major objectives: 1) the development of a digital and multiplex assay (Digiplex), where each miRNA is accurately quantified independently at the single molecule level using a target-specific molecular program; 2) the exploration of molecular neural networks (MolNNet) for direct recognition of concentration patterns. This objective involves sophisticated molecular program architectures that take as input multiple miRNA concentrations, then carry out signal processing and report the sample type (e.g. “healthy” or “diseased”), all these reactions happening simultaneously in a single tube.
This project is highly interdisciplinary, gathering expertise in DNA nanotechnology, microfluidics, surface chemistry and machine learning. The expected outcomes include the advance of cutting-edge miRNA quantification technologies combining single-molecule amplification with a multiplex readout (up to 100 targets). In the long term, the exploration of in moleculo neural networks is foreseen to trigger a whole new field of research, providing a ground-breaking approach for molecular diagnostics.
MoP-MiP will develop novel approaches for miRNA profiling building on the latest advances in molecular programming (MP). MP deals with the design of artificial DNA reaction networks capable of information processing, which will be exploited throughout this project to build smarter biosensors. The team will pursue two major objectives: 1) the development of a digital and multiplex assay (Digiplex), where each miRNA is accurately quantified independently at the single molecule level using a target-specific molecular program; 2) the exploration of molecular neural networks (MolNNet) for direct recognition of concentration patterns. This objective involves sophisticated molecular program architectures that take as input multiple miRNA concentrations, then carry out signal processing and report the sample type (e.g. “healthy” or “diseased”), all these reactions happening simultaneously in a single tube.
This project is highly interdisciplinary, gathering expertise in DNA nanotechnology, microfluidics, surface chemistry and machine learning. The expected outcomes include the advance of cutting-edge miRNA quantification technologies combining single-molecule amplification with a multiplex readout (up to 100 targets). In the long term, the exploration of in moleculo neural networks is foreseen to trigger a whole new field of research, providing a ground-breaking approach for molecular diagnostics.
Objective 1: we made a proof of principle using microscopic beads, which are decorated with DNA strands for 1) identifying the target miRNA by a combination of fluorescent bead-code, 2) capturing miRNA from the samples and 3) report its presence through the emission of a fluorescence signal. Importantly, because the number of beads is higher than the number of miRNAs, each bead captures either 0 or 1 miRNA, this according to a random (Poisson) distribution. The use of magnetic beads allows to wash them, and therefore to discard all components of the samples that can interfere with the subsequent amplification step.
The particles are then isolated in water-in-oil droplets (using a microfluidic chip, 1 bead per drop), together with an amplification mixture composed of DNA strands and enzymes that transform the miRNA target into a fluorescence signal. Therefore, the particles having captured their cognate miRNA would trigger the amplification reaction and in turn activate the reporting DNA strand (resulting in a positive fluorescence signal on the particle). By contrast, the particles that have not captured their target miRNA would stay off (negative particles). All particles are finally analyzed by a flow cytometer, an instrument that measures the fluorescence intensity of individual particles as they pass through a laser beam. For each particle is measured the fluorescence intensity of its bead code (which serves as index the particle to its target miRNA) and of the reporting strand (to classify the particle as positive or negative). The particle ON/OFF ratio allows to compute using statistics, for each target miRNA, its concentration in the sample. We have made a proof of concept for the digital and multiplex quantification of 6-10 miRNAs. The new methodology has been optimized to detect endogenous miRNA from cell extract. We have adapted this technique to the detection of other biological targets, in particular enzymes (12 demonstrated, e.g. DNA polymerases, nucleases such as Cas9, glycosylases, kinases, phosphatases…). We showed that beyond its use for biosensing applications, the assays can be used to investigate the functional heterogeneity of a population of enzymes, of great interest for directed evolution of proteins.
The second objective of this work is being conducted in collaboration with the group of Anthony Genot (LIMMS CNRS/U. of Tokyo), specialized DNA computing. The idea is to create DNA/enzyme reaction networks that mimic the architecture of an artificial neural network (ANN). Like an ANN that classify pictures of cats and dogs, our molecular neural network aims at classifying patient samples according to their content in miRNA. We have shown that our molecular toolbox can implement the essential ingredient of such DNA neuron that are: i) tunable weights: this is achieved by adjusting the concentration of the DNA template that convert the input molecule into a signal strand. We also show that negative weight can be implement by having the corresponding DNA template produce an anti-signal strand. ii) summation: all signal strands produced from different inputs can be collected by an amplification template. iii) nonlinear activation function: Once a threshold (set by a drain template), is exceeded, i.e. the signal strands reaches a given concentration, the amplification template exponential amplify the signal strand, resulting in a sharp transition from low to high concentration of the signal strand. Such simple DNA neuron can be used to build a perceptron-like network, allowing for the linear classification of samples based on up to 10 miRNA input. We demonstrated linear as well as nonlinear space partitioning. This is the culminating demonstration of an article published in Nature in 2022.
The particles are then isolated in water-in-oil droplets (using a microfluidic chip, 1 bead per drop), together with an amplification mixture composed of DNA strands and enzymes that transform the miRNA target into a fluorescence signal. Therefore, the particles having captured their cognate miRNA would trigger the amplification reaction and in turn activate the reporting DNA strand (resulting in a positive fluorescence signal on the particle). By contrast, the particles that have not captured their target miRNA would stay off (negative particles). All particles are finally analyzed by a flow cytometer, an instrument that measures the fluorescence intensity of individual particles as they pass through a laser beam. For each particle is measured the fluorescence intensity of its bead code (which serves as index the particle to its target miRNA) and of the reporting strand (to classify the particle as positive or negative). The particle ON/OFF ratio allows to compute using statistics, for each target miRNA, its concentration in the sample. We have made a proof of concept for the digital and multiplex quantification of 6-10 miRNAs. The new methodology has been optimized to detect endogenous miRNA from cell extract. We have adapted this technique to the detection of other biological targets, in particular enzymes (12 demonstrated, e.g. DNA polymerases, nucleases such as Cas9, glycosylases, kinases, phosphatases…). We showed that beyond its use for biosensing applications, the assays can be used to investigate the functional heterogeneity of a population of enzymes, of great interest for directed evolution of proteins.
The second objective of this work is being conducted in collaboration with the group of Anthony Genot (LIMMS CNRS/U. of Tokyo), specialized DNA computing. The idea is to create DNA/enzyme reaction networks that mimic the architecture of an artificial neural network (ANN). Like an ANN that classify pictures of cats and dogs, our molecular neural network aims at classifying patient samples according to their content in miRNA. We have shown that our molecular toolbox can implement the essential ingredient of such DNA neuron that are: i) tunable weights: this is achieved by adjusting the concentration of the DNA template that convert the input molecule into a signal strand. We also show that negative weight can be implement by having the corresponding DNA template produce an anti-signal strand. ii) summation: all signal strands produced from different inputs can be collected by an amplification template. iii) nonlinear activation function: Once a threshold (set by a drain template), is exceeded, i.e. the signal strands reaches a given concentration, the amplification template exponential amplify the signal strand, resulting in a sharp transition from low to high concentration of the signal strand. Such simple DNA neuron can be used to build a perceptron-like network, allowing for the linear classification of samples based on up to 10 miRNA input. We demonstrated linear as well as nonlinear space partitioning. This is the culminating demonstration of an article published in Nature in 2022.
Objective 1: Digital droplet PCR has become a reference technique for the accurate quantification of nucleic acids. However, the multiplexing capabilities is currently limited to 3-6 targets. Our Digiplex assay, although less mature and robust, has the potential to dramatically extend this limit. We have now demonstrated the detection of 10 targets simultaneously and intend to show 100 targets in the future.
Objective 2: Neuromorphic computing was explored in the past by famous groups that have pioneered DNA nanotechnology (notably E. Winfree, L. Qian, G. Seelig). The neural architectures these groups proposed contain a single neuron layer, hence limiting the classification to linearly separable data (e.g. setting a separatrix line for 2D data). In the article we published in Nature, one essential breakthrough was the construction of multilayer architectures for solving nonlinear classification problem (like a decision tree in diagnostic). We applied this multilayer architecture to the nonlinear partitioning of a 2D space composed of 2 miRNA concentrations. The second significant advancement in this work was the experimental demonstration of a 10-bit majority voting algorithm. In such algorithm, all DNA strand inputs (voters) have the same weight (experimentally adjusted) and the network reports a positive signal only of the sum of positive voter exceeds a threshold (set to 5, in a 10-input system). The classification exceeds 90 % accuracy and veto function was implemented in a more sophisticated version of the algorithm. Note that such majority voting algorithm was limited to 3-input in the prior art.
Objective 2: Neuromorphic computing was explored in the past by famous groups that have pioneered DNA nanotechnology (notably E. Winfree, L. Qian, G. Seelig). The neural architectures these groups proposed contain a single neuron layer, hence limiting the classification to linearly separable data (e.g. setting a separatrix line for 2D data). In the article we published in Nature, one essential breakthrough was the construction of multilayer architectures for solving nonlinear classification problem (like a decision tree in diagnostic). We applied this multilayer architecture to the nonlinear partitioning of a 2D space composed of 2 miRNA concentrations. The second significant advancement in this work was the experimental demonstration of a 10-bit majority voting algorithm. In such algorithm, all DNA strand inputs (voters) have the same weight (experimentally adjusted) and the network reports a positive signal only of the sum of positive voter exceeds a threshold (set to 5, in a 10-input system). The classification exceeds 90 % accuracy and veto function was implemented in a more sophisticated version of the algorithm. Note that such majority voting algorithm was limited to 3-input in the prior art.